3

I've been working on a guitar tuner Java application for quite some time and have finally managed to get the pitch (fundamental frequency) detection to accurately determine the fundamental frequency of the input using an FFT and the Harmonic Product Spectrum. This works perfectly when I use my MacBook Pro's default microphone, despite the fact that the signal is a little noisy, but I would ideally like to use my external USB audio sound interface (Lexicon Lambda) with an electric guitar directly plugged in. I have tried this with the exact same code that works perfectly for the mic input, but the pitch detected is different. Here are my results of fundamentals detected for each string (the guitar is in tune already).

Sample rate = 8000
Samples taken = 16000
Frequency resolution = 0.5Hz

E - Mic: 82.5Hz - Lexicon Lambda: 124.5Hz
A - Mic: 110.0Hz - Lexicon Lambda: 165.5Hz
D - Mic: 147.0Hz - Lexicon Lambda: 221.5Hz
G - Mic: 196.0Hz - Lexicon Lambda: 293.5Hz
B - Mic: 247.0Hz - Lexicon Lambda: 371.0Hz
E - Mic: 329.5Hz - Lexicon Lambda: 495.0Hz

I have messed around a bit with these numbers and it seems that the result found using the Lexicon Lambda is generally about 1.5 times the size of the accurate result. But rather than just adding this extra calculation to get to the right number I would like to make it so that both methods of obtaining audio data return accurate results.

Does anyone have any idea what the cause of this inconsistency might be?

Edit

So I am pretty sure that the problem is due to the Lambda using a bit-depth of 24 while my program (and the mic input) works with a bit-depth of 16. Rather than just changing it to deal with 24-bit data, does anyone know of a way to detect the available bit-depth of the current audio device?

Using the Lambda is the priority, since the signal is less noisy; however, I have just been trying to get things to work with a bit depth of 24 specified, and I get the IllegalArgumentException "No line matching interface TargetDataLine supporting format PCM_SIGNED 8000.0 Hz, 24 bit, mono, 3 bytes/frame, little-endian is supported." I tried using 44.1kHz as well and I get the same error, which suggests the Lambda isn't actually working with 24-bit audio or perhaps there is another issue. Does anyone have any experience working with 24-bit audio in Java Sound? It seems like there might be more to consider than just changing the bit-depth in order to make this work.

Thanks

nihilo90
  • 109
  • 1
  • 9
  • *"exact same code"* What code? For better help sooner, post an [SSCCE](http://sscce.org/). – Andrew Thompson Feb 10 '12 at 15:35
  • I have not tried this, but are you sure your Lamdba driver supports 8kHz sampling rate? Pretty uncommon actually – guitarflow Feb 10 '12 at 15:39
  • +1 @nihilo interesting project – therobyouknow Feb 10 '12 at 16:15
  • +1 As myself and @guitarflow have stated, the fact that your input sources have a bit depth of 24 bits and 16 bits and that this ratio ( 1.5 or 66% depending on which way around you look at it) is the same as the ratio of the frequencies gives some clue as to where the problem lies. If your calculation of frequency uses filesize somehow then the 24/16 = 1.5 or 16/24=1.5 theory is the root cause. – therobyouknow Feb 13 '12 at 14:39
  • Did you find the root cause yet? – therobyouknow Mar 06 '12 at 16:38
  • Well it certainly was the bit depth. I haven't actually made that much progress in terms of using 24-bit samples, but I discovered the setting to change my audio interface to use 16-bit. With 16-bit sampling selected, my application works just fine :) – nihilo90 Mar 09 '12 at 13:10

3 Answers3

3

The mic values are 66% of the Lexicon Lambda, e.g. E - Mic: 329.5Hz divided by Lexicon Lambda: 495.0Hz = about 0.66. Wonder if that is significant?

Is the sampling rate from the mic and the rate from the Lexicon Lambda the same?

Is the bit depth the same? I wonder if they are different, e.g. one is 16bit and the other 24 bit? That might explain the 66% value in the calculations some how...

therobyouknow
  • 6,604
  • 13
  • 56
  • 73
2

1st idea

The sample bitrate could be the key!! A professional audio interface will by default use 24 bits, 16 bits is an 80's standard that is still used a lot in consumer electronics.

If your FFT expects 16 bit values and you pass 24 bit, your sound will be "stretched". ( Imagine a sine wave to make it clearer )

This explains the frequency drift perfectly!! So you should either convert the 24 bits to 16 or rewrite the FFT.

2nd idea

I found out on the Lexikon website that the interface just supports 44.1 kHz and 48 kHz. So I guess the Lambda is by default set to 48 kHz. If you request an 8 kHz sample rate, the interface can't handle that and stays at 48 kHz I suppose.

So just try a common samplerate on both devices and compare again.

guitarflow
  • 2,930
  • 25
  • 38
  • 1
    I tried with 44.1kHz for both and the results are the same. I'll have a go at using 24-bit samples instead though. The fact that 24 = 1.5 * 16 and the results from the Lambda are 1.5* the accurate readings is a pretty big clue. I think you were right before the edit. – nihilo90 Feb 10 '12 at 17:26
  • @nihilo90 Ok,edited. Would be interested once you get it figured out. – guitarflow Feb 10 '12 at 19:03
0

I had a sweet theory for you but on further thought it was complete bs. In other news, I just wanted to mention that for frequency detection, the precision of your FFT is far more important than the precision or noise on the microphone samples. Quantization noise on the mic will add noise all over the spectrum and not influence the FFT peaks much, and background noise will also hopefully lie far below the detected FFT peak and not influence the result much.

I have a tuner app on android and I messed around with fixed point optimizations in the FFT but the results were horrible. On the other hand, the app works great even if you're far away from the phone and on crappy mic hardware, so I would think you don't really need all 24 bits from your Lexicon, just ensure you use as much resolution as possible (double, preferably) for the FFT multiplications and additions.

c.fogelklou
  • 1,781
  • 1
  • 13
  • 26
  • Using 16 bits is good enough, but it would seem that the Lexicon Lambda doesn't work with 16 bit samples. That's the only reason I'm trying to get it to work with 24 bit samples really. No success so far :( – nihilo90 Feb 12 '12 at 22:56